MTSCNet first decomposes long-term time series data into multiple scales through multi-scale average downsampling, thus disentangling complex temporal patterns. Subsequently, to address the non-stationarity and signal entanglement of time series data, a frequency-domain selective decomposition method is employed to obtain seasonal and trend components, which are modeled separately at each scale. The seasonal component is processed through bottom-up MLP mixing, progressively compressing high-frequency seasonal details and aggregating them into lower frequencies, thus extracting stable high-frequency seasonal patterns and suppressing noise. The trend component is processed through top-down Chebyshev-KAN mixing, progressively propagating low-frequency global trends to higher frequencies and enhancing nonlinear representations. Meanwhile, normalized time indices are introduced as explicit independent temporal variables, enabling the model to more accurately characterize trend evolution over time and improve long-term forecasting and extrapolation capabilities. After recombining the seasonal and trend components, an SE-attention channel mechanism is introduced into the hybrid representation, performing adaptive recalibration of channel features to highlight key variables and suppress noise interference. Extensive experimental results on six real-world time series datasets demonstrate that MTSCNet achieves state-of-the-art performance.